SEO vs GEO and AEO: why AI visibility does not follow your rankings

SEO helps AI search visibility, but it is not the same thing. In Ahrefs’ study of 75,000 brands, about one in four that passed the study’s search thresholds had zero AI Overview mentions, and backlinks were the weakest of the seven signals it measured. The most costly mistake we encounter is treating GEO and AEO as nothing more than a new label for an existing search program.

Two documents published this year offer clashing guidance on AI visibility, and each deserves attention. Google’s AI optimization guide argues there is no separate optimization path for AI, since generative features rely on the same ranking machinery as traditional search. By contrast, Jessica Bowman’s article in Search Engine Land maintains that search engine optimization gets you only halfway there, and the part that determines whether an assistant recommends you may live entirely outside marketing. Both cannot fully explain how generative engine optimization works. Because we do this every day, we looked for where the evidence actually points.

What Google actually claims

Google’s stance on generative engine optimization is strikingly direct. It says AI Overviews and AI Mode are powered by core Search ranking and quality systems, which means ordinary search engine optimization covers the job. The guide then enumerates what it says makes no difference, and the list is longer than many expect. Google Search does not use llms.txt or other AI text files. Generative AI search does not require structured data. Breaking content into tiny chunks is unnecessary. Reworking pages specifically for language models is pointless, because the systems already understand synonyms and context. And trying to manufacture product mentions around the web is not, according to Google, an effective AI visibility tactic.

Taken literally, that guidance eliminates most of what the answer engine optimization market sells. If your technical SEO is solid, you publish useful content, and your site remains crawlable, Google’s view is that you are finished.

There is one limit, though. That page applies only to Google’s own answer experiences, and nothing beyond them. AI Overviews and AI Mode are just two of the eight generative engines a buyer may use. ChatGPT, Claude, Perplexity, Gemini, Grok and ChatGPT Shopping all retrieve, rank and cite according to their own rules, and this guidance does not establish how those other experiences select sources.

What Bowman argues instead

Bowman begins from a blind spot that most search engine optimization frameworks do not name. Being mentioned by an AI assistant is not the same thing as being recommended by it. A brand can have clean technical SEO, genuinely strong content and regular citations, yet still disappear the moment the assistant stops summarizing a topic and starts advising a purchase.

Her point is uncomfortable. Sometimes the AI recommendation is lost because the model accurately understood your product. It noticed the maintenance load, the missing integration, the support complaints, or the capability you simply do not offer. No amount of content shaped like an answer can repair a product that is truly behind, and no GEO consultant can write around that. The people who can fix it sit in Product, Support, Finance and Design. In that framing, the search team’s real role is diagnosis plus internal persuasion: they own the program, while other teams own the remedy.

Google’s guide has no bucket for that scenario. Every fix it mentions is either a content fix or a crawling fix. The case where the system is simply correct about you is absent.

Placed next to each other, the two views barely overlap:

The data does not support either one cleanly

A third source brings more evidence into the discussion. Ahrefs studied 75,000 brands and examined which signals correlate with inclusion in AI Overviews. If Google’s framing were the whole story, the classic search engine optimization metrics should dominate. They do not.

Backlinks ranked last. In this dataset, backlink counts sit at the bottom, at about one third the correlation of branded mentions on other sites. Domain Rating, the metric agencies place on the first slide of nearly every pitch, lands in the middle. What leads is being named in text on sites you do not control.

That finding pushes back on Google and on conventional search optimization at the same time. It weakens the link-first doctrine inherited by AEO from SEO. It also sits awkwardly beside Google’s statement that pursuing manufactured product mentions across the web is ineffective, because unlinked brand mentions are the strongest measured signal. The two ideas can be reconciled, since Google is warning against manufactured mentions and the study measured organic ones, but the practical advice a marketer takes away from each is almost the opposite.

Still, keep the Ahrefs caveat in mind. The authors explicitly say correlation is not causation, and every coefficient they found was only moderate at best. Brands with more mentions are also usually larger, older and better funded. The mechanism has not been isolated.

Three more places the advice contradicts itself

The clearest example is llms.txt. Google says its Search systems ignore the file entirely. Meanwhile, much of the generative engine optimization industry still advises publishing one, and the agent tooling world has adopted it for an entirely different purpose. Both statements can be true because they refer to different systems. A file that does nothing for Google AI Overviews may still be useful in the agentic layer, where assistants pull context to complete tasks. The contradiction only appears when you stop asking which engine is being discussed.

Structured data follows the same pattern. Google says schema markup is not required for its generative AI search features. Many AEO practitioners, however, treat JSON-LD as a foundation. Note the wording: not required is not the same as useless. Machine readability and ranking necessity are separate questions, and the guide answers only one of them.

Recency is another split. Classic SEO tends to reward evergreen content that builds authority over years. Answer engines, by contrast, lean strongly toward fresh material, and a page that was being cited constantly in spring can quietly stop being cited by autumn without losing a single ranking position. If your GEO reporting is just a rank tracker wearing a new label, you will miss that change.

Why the engines disagree with each other too

Beneath all of this is a basic reason universal AI search advice is shaky: the engines are not built the same way. Google AI Overviews draw from the Google index. ChatGPT’s search relies on a different index. Perplexity uses its own live retrieval. Claude, Gemini and Grok each layer their own reranking and citation behavior on top.

That means the same change can produce different outcomes depending on the surface, and any single “AI visibility” number is really an average across systems that do not agree. This is why we sample repeatedly across engines instead of checking once, and why our visibility scores include confidence bands rather than a neat single figure. A one-number AI visibility score with no uncertainty attached is a marketing artifact, not a measurement.

The proof that crawlable does not mean recommended

The strongest evidence that AI visibility is its own discipline comes from that same Ahrefs dataset. To be included, a brand had to have a Domain Rating above 40 and at least one keyword with 800 monthly searches or more. These are not obscure websites. They are businesses that already succeeded at search engine optimization.

About a quarter of them were mentioned zero times in AI Overviews.

Roughly one in four brands in this filtered sample had no AI Overview mentions at all, even though they met the study’s domain and search-volume thresholds. Those thresholds do not prove that every aspect of their SEO is working, and the missing mentions do not establish why they are absent. If AI search visibility followed rankings, that group should be nearly empty.

Most teams stop at crawl data, so it is worth stating exactly what it proves. If GPTBot, ClaudeBot and Googlebot fetch your pages and receive a 200, that only means the assistants can read your pages. It tells you nothing about whether they recommend you. You need both measurements, and confusing them is how brands end up reporting healthy AI crawler traffic while losing every buyer question in their category.

The more useful question is which sources the answer engines actually quote in your category, because that is where the correlation points. If assistants keep citing three review platforms and a community thread when they answer your buyers’ questions, those four pages become your GEO roadmap. It is not glamorous, and it looks more like public relations and partnerships than technical SEO.

Which GEO advice is backed by evidence, and which is speculation?

Sort any tip by what stands behind it.

Documented by Google, for Google’s own AI features only: AI Overviews and AI Mode use the core Search ranking and quality systems; Google Search does not use llms.txt; structured data is not required for its generative features.

Measured, but only as correlation: in Ahrefs’ 75,000-brand study, branded web mentions had the strongest link to AI Overview mentions (0.664) and backlink counts the weakest (0.218). The authors say plainly that correlation is not causation.

Speculation: any claim that one tactic causes ChatGPT, Perplexity or Claude to recommend you. No source in this article shows that for an engine other than Google, and that covers llms.txt for ChatGPT and schema markup as a ranking factor.

If a tip does not say which engine it applies to, file it under speculation until it does.

So what should you actually do

Use the evidence to separate access, content and product work.

Do the search engine optimization work, because it is necessary, inexpensive and foundational for both GEO and AEO. Keep the site crawlable, keep it fast, publish pages that answer real buyer questions clearly, and make sure AI crawlers can reach you. If an assistant cannot read your pages, it cannot recommend you. Google is right that this is the base layer and right that most AI-specific file tricks are a waste of time.

But do not treat that as sufficient, because the evidence says it is not. Invest in being named on the pages assistants actually cite for your category, which usually means review platforms, comparison roundups, community threads and coverage from writers other people trust. That is where the strongest measured correlation lives, and it is ordinary earned marketing rather than a technical hack. Our guide on getting cited by AI explains the mechanics, and AEO agency vs AI visibility software covers whether to hire help for it.

And when the measurement keeps saying you are losing a question you ought to win, take Bowman seriously and ask whether the answer engines are correct about your product. That conversation belongs with Product or Support, not in your content calendar.

Most of all, measure by engine and measure repeatedly. Every apparent contradiction in this article disappears once you stop asking what works for AI in general and start asking what works on which surface, for which buyer question. A rank tracker cannot tell you that, and neither can vendor guidance, including this page.

Geofound brings these investigations into one workflow: daily AI tracking, competitor SEO and paid ads research, and recommendations with playbooks. Use rankings and keyword opportunities to investigate search coverage, backlinks to find relevant sources, and competing ads to compare offers. Review AI mentions and citations alongside that research to decide which page or message needs work.

Begin with the free scan: three buyer questions, one engine, and the names it returned. Explore Geofound's plans for ongoing tracking across up to eight AI surfaces, SEO and paid ads research, and help acting on the findings. You can also compare the alternatives.

FAQ

What is the difference between AI visibility and SEO? SEO measures whether your pages rank in search results. AI visibility measures whether assistants such as ChatGPT, Perplexity and Google AI Overviews name your brand when a buyer asks a question. The two overlap, because AI Overviews draw on Google’s index and every assistant needs pages it can read. They also split apart: in the Ahrefs study of 75,000 brands, about one in four brands that passed the study’s search thresholds had zero AI Overview mentions.

Is GEO just SEO with a new name? No. They share a base but diverge above it. Search engine optimization focuses on your pages so they rank for queries. Generative engine optimization focuses on your presence across the sources assistants cite so you are recommended in answers. The Ahrefs correlation data is the clearest proof they are not the same discipline, since backlinks were the weakest measured signal and unlinked brand mentions the strongest.

What is the difference between GEO and AEO? In practice, people use the terms interchangeably. Answer engine optimization usually stresses being the quoted answer, while generative engine optimization usually stresses being named in a synthesized recommendation. Both describe the same aim: getting your brand into AI answers. Neither is a Google-recognized category.

Should I publish an llms.txt file? Not for Google AI Overviews, which ignore it completely. It may be useful in agent tooling settings. Do not expect it to change whether ChatGPT or Perplexity recommends you, and do not let it become your entire GEO strategy.

Does structured data help AI visibility? Google says it is not required for its generative features, which is not the same as saying it is useless. Schema is inexpensive, it makes your product machine readable, and it offers other benefits. Just do not assume it will be the reason an assistant starts recommending you.

How often should I check AI visibility? For most brands, weekly is enough, and the trend tells you more than any single day. Answers change from run to run, so one check tells you very little. That variation is exactly why AI visibility measurement needs repeated sampling.